Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images

Page created by Juan Castro
 
CONTINUE READING
Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images
remote sensing
Article
Cross-Domain Scene Classification Based on a Spatial
Generalized Neural Architecture Search for High Spatial
Resolution Remote Sensing Images
Yuling Chen, Wentao Teng, Zhen Li, Qiqi Zhu * and Qingfeng Guan

 School of Geography and Information Engineering, China University of Geosciences, Wuhan 430078, China;
 ylchennn@163.com (Y.C.); 20171003080@cug.edu.cn (W.T.); zhen_li@cug.edu.cn (Z.L.); guanqf@cug.edu.cn (Q.G.)
 * Correspondence: zhuqq@cug.edu.cn

 Abstract: By labelling high spatial resolution (HSR) images with specific semantic classes according
 to geographical properties, scene classification has been proven to be an effective method for HSR
 remote sensing image semantic interpretation. Deep learning is widely applied in HSR remote
 sensing scene classification. Most of the scene classification methods based on deep learning assume
 that the training datasets and the test datasets come from the same datasets or obey similar feature
 distributions. However, in practical application scenarios, it is difficult to guarantee this assumption.
 For new datasets, it is time-consuming and labor-intensive to repeat data annotation and network
 design. The neural architecture search (NAS) can automate the process of redesigning the baseline
 network. However, traditional NAS lacks the generalization ability to different settings and tasks.
 In this paper, a novel neural network search architecture framework—the spatial generalization
 
  neural architecture search (SGNAS) framework—is proposed. This model applies the NAS of spatial
 generalization to cross-domain scene classification of HSR images to bridge the domain gap. The
Citation: Chen, Y.; Teng, W.; Li, Z.;
Zhu, Q.; Guan, Q. Cross-Domain
 proposed SGNAS can automatically search the architecture suitable for HSR image scene classification
Scene Classification Based on a and possesses network design principles similar to the manually designed networks. To obtain a
Spatial Generalized Neural simple and low-dimensional search space, the traditional NAS search space was optimized and
Architecture Search for High Spatial the human-the-loop method was used. To extend the optimized search space to different tasks, the
Resolution Remote Sensing Images. search space was generalized. The experimental results demonstrate that the network searched by
Remote Sens. 2021, 13, 3460. https:// the SGNAS framework with good generalization ability displays its effectiveness for cross-domain
doi.org/10.3390/rs13173460 scene classification of HSR images, both in accuracy and time efficiency.

Academic Editor: Filiberto Pla Keywords: high spatial resolution images; cross-domain scene classification; neural architecture
 search; spatial generalization; deep learning
Received: 20 July 2021
Accepted: 25 August 2021
Published: 1 September 2021

 1. Introduction
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
 With the continuous development of satellite sensors, the resolution of remote sensing
published maps and institutional affil- images is improving, and fine-scale information can be obtained from high spatial resolu-
iations. tion (HSR) remote sensing images [1]. However, HSR remote sensing images have many
 textural, structural, and spectral characteristics [2,3]. These data demonstrate the phenom-
 ena of a complex spatial arrangement with high intraclass and low interclass variabilities,
 giving rise to difficulties in image classification and recognition [4,5]. Pixel-based remote
Copyright: © 2021 by the authors.
 sensing image classification methods frequently consider that the same kinds of ground ob-
Licensee MDPI, Basel, Switzerland.
 jects have the same spectral characteristics, while different ground objects in spectral space
This article is an open access article
 use separability as the classification premise. Therefore, the pixel-level remote sensing
distributed under the terms and image classification method cannot be effective. The object-oriented classification method is
conditions of the Creative Commons proposed and widely used in HSR images [6–8]. The object-oriented classification method
Attribution (CC BY) license (https:// can accurately identify the target information and features in HSR remote sensing images.
creativecommons.org/licenses/by/ However, the difference between the underlying features and the high-level semantic
4.0/). information still makes the traditional object-oriented image interpretation method unable

Remote Sens. 2021, 13, 3460. https://doi.org/10.3390/rs13173460 https://www.mdpi.com/journal/remotesensing
Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images
Remote Sens. 2021, 13, 3460 2 of 22

 to obtain the complex semantic information of the important areas in HSR remote sensing
 images [9]. Eliminating the semantic-gap [10] between low-level features and high-level
 semantic information to acquire scene semantic information of HSR images has become
 one of the hot-spots for remote sensing image processing and analysis [11,12].
 Scene classification enables obtaining high-level semantic information about different
 scenes by automatically labelling HSR remote sensing images to establish the relationship
 between the underlying features and the high-level semantic information [13]. The key
 to scene classification lies in the extraction of essential features from the images. In
 recent years, the scene classification method without considering the object prior has
 been widely used in HSR remote sensing images and can be divided into three main
 types [14,15]. (1) Scene classification methods based on low-level features. This kind
 of method extracts the bottom features from the HSR images such as the color, texture,
 and shape attributes of the image, and uses classifiers to identify scenes. The bottom
 features can be obtained directly from images without external knowledge including
 global features [16] and local features [17]. For example, Yang and Newsam compared
 the scale invariant feature transformation (SIFT) [18] and Gabor texture features. Dos
 Santos et al. tested a variety of global color and texture features such as color histogram
 (CH) [19] and local binary pattern (LBP) [20]. However, this method has difficulty in
 describing the complex spectral and spatial characteristics of ground objects in scenes, thus
 the classification results are often poor [16,21]. (2) Scene classification methods based on
 mid-level features. By extracting the local features of the scene, this method maps the local
 features to the dictionary or parameter space to obtain the mid-level features with stronger
 discrimination. Then, the mid-level features are input into the classifier to obtain the scene
 labels [22]. This kind of method mainly includes bag-of-visual-words (BOVW), feature
 coding, and probabilistic topic models [22–25]. Luo et al. [23] constructed descriptors for
 satellite image by connecting color and texture features, and quantified the features of
 all image blocks into several visual words by K-means clustering. However, the scene
 classification method based on mid-level features often ignores the spatial distribution
 between features and lacks the transferability between domains, which greatly limit the
 expression effectiveness and model universality of HSR image scene classification. (3) Scene
 classification methods based on high-level features, which automatically extract features
 from images through training a deep network. Deep learning methods can mainly be
 divided into two types: supervised feature learning and unsupervised feature learning
 based methods [26]. With the generalization ability of the deep learning model and the
 similarity of data between different fields [27–29], it is possible to transfer network weight
 from one domain to another, which is beneficial to network initialization and saves in the
 time consumption of training [27].
 The earliest deep learning method introduced into HSR remote sensing images is
 the unsupervised feature learning method based on a significant sample selection strat-
 egy [30]. The training data of this method is unlabelled. The deep learning method based
 on supervised features uses labelled image data to train and optimize neural networks
 to classify images. The CNN is one of the most widely used models in deep learning
 based on supervised features. According to the development process and direction of CNN
 scene classification, traditional CNN scene classification methods can be divided into three
 types [31]: (1) Classical convolutional neural network, for example, AlexNet [32] is com-
 posed of a convolution layer, a fully connected layer, and a softmax classifier. VGGNet [33]
 inherits the framework of LeNet [34] and AlexNet, increasing the depth of the network and
 obtaining more effective information. These CNN networks have achieved good results in
 scene classification fields. (2) Convolutional neural network based on the attention mecha-
 nism, which means that the neural network uses the attention mechanism to emphasize
 the useful part more, while inhibiting the less useful part [35–37]. (3) Lightweight network.
 One of the classic models is SqueezeNet [38], which achieves the same accuracy as AlexNet.
 However, the parameters of SqueezeNet are only one-fiftieth of the parameters of AlexNet.
 Deep learning, particularly convolutional neural networks (CNNs), has quickly become a
Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images
Remote Sens. 2021, 13, 3460 3 of 22

 popular topic in scene classification applications [39–42]. In the continuous development of
 deep learning, a network design principle has been developed for manual design networks.
 The outcome of this design principle is not only suitable for a single network, but can also
 be generalized and applied to different settings and tasks.
 However, most deep learning-based scene classification methods usually assume
 that the training and testing data are the same datasets or share the same distribution. In
 practical applications, since the training and testing data often come from different regions
 or sensors, the feature distributions are quite different. This phenomenon is referred to as
 the data shift, and it makes the above assumptions unreliable [43,44]. When faced with
 new data, manually designed networks often need to once again perform data annotation
 and network design [43,45,46]. However, data annotation is time-consuming and labor-
 intensive. Manually designed networks have difficulty adapting well to new datasets
 or tasks due to insufficient experiments or lack of experience [47]. Inheriting from the
 architecture of natural image recognition to design a new network is another method in
 the field of manual design networks [47]. However, the design of the architecture for this
 method often ignores the characteristics of data without considering the complexity and
 specificity of HSR images.
 A neural architecture search (NAS) method that uses neural networks to automatically
 search neural network structures [48–53] has been proposed. The network searched by
 the NAS has been applied to large-scale image classification, semantic segmentation,
 recognition tasks, and scene classification. For example, Barret Zoph et al. [54] proposed
 a NAS based on reinforcement learning, which used a recursive network to design the
 network and train the network. The experimental results of the Penn Treebank (PTB)
 language modelling experiments showed that the model was superior to other advanced
 models on the PTB dataset. Li et al. [53] proposed the Auto-DeepLab model and studied
 the semantic segmentation method based on the NAS. Compared to manually designed
 networks, NASs have the advantages of automatically searching network architectures
 with high efficiency. However, the traditional NAS has the problem of poor generalization
 ability and high resource consumption in the search for architectures [55]. The search
 result of the traditional NAS is to tune a single network instance to a specific setting. The
 design paradigm of traditional NAS has limitations, which cannot help to discover the
 general design principles of the network and extend them to different settings [56]. Based
 on the above problems, He et al. [56] proposed RegNet and found a new general design
 paradigm of NAS, and the network obtained by RegNet can be generalized in various
 settings. Under comparable training settings and flops, the RegNet models outperformed
 the popular EfficientNet models while being up to 5× faster on GPUs. Similar to the
 manually designed CNN, the traditional NAS often assumes that the features of the
 training and testing datasets are the same or similar in scene classification. Cross-domain
 scene classification refers to scene classification tasks in which training sets and test sets
 come from different distributions. and can help to ameliorate the effect of the data shift [57].
 This helps to optimize the model to meet the requirements of practical scene classification
 application [58,59]. The generalization ability of the model, which refers to the ability
 of the learned model to predict an unknown dataset, can be evaluated by the results of
 cross-domain scene classification [44,60,61].
 To solve the problem that manual network design is time-consuming and difficult,
 and the generalization ability of the traditional NAS is poor, a novel neural network search
 framework—the spatial generalized neural architecture search (SGNAS) framework—was
 proposed in this paper for HSR image cross-domain scene classification. The SGNAS
 focuses on the simplification and design of the search space. Within the SGNAS frame-
 work, the purpose of this study was to design a simple search space with high efficiency
 and generalization. Based on the design of the NAS, the setting of its search space was
 optimized in this study. After optimization, the human-the-loop method was applied to
 obtain a low-dimensional search space composed of simple networks [56]. To make the
 search space simpler than the original search space, which essentially aims to improve the
Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images
Remote Sens. 2021, 13, 3460 4 of 22

 search efficiency, a low-computation, low-epoch training regime was used. To ensure that
 the SGNAS can be applied to different settings and tasks, this study combines the general-
 ization of manual network design principles in the preoptimized and trained search space
 of the NAS. Based on the designed search space, the model search method integrating the
 random search strategy with the performance estimation method of training from scratch
 was used to search the network. Finally, the final model was applied to cross-domain scene
 classification to verify the generalization ability of the model.
 The major contributions of this paper are as follows:
 The SGNAS framework was proposed to discover discriminative information from
 HSR imagery for cross-domain scene classification. Based on the level of search space,
 the semiautomatic NAS combines the advantages of manual network design, and the
 traditional NAS was designed. We designed the search space in a low-computation, low-
 epoch regime, which can be generalized to heavier computation regimes, schedule lengths,
 and network block types. SGNAS overcomes the bottleneck of the fixed design paradigm
 of a traditional manual design network and the difficulty of network redesign.
 The network searched by SGNAS implements cross-domain scene classification tasks
 between different datasets in an unsupervised way. In other words, the training and testing
 datasets will be two groups of different datasets with large differences in characteristics.
 To search for suitable models for cross-domain scene classification, the evaluation feedback
 in the performance evaluation was obtained by the cross-domain scene classification in
 this study.
 The rest of this paper is organized as follows. Section 2 introduces the materials and
 provides a detailed description of the construction of the proposed SGNAS framework.
 Section 3 introduces the experimental results of the proposed scene classification method.
 Section 4 discusses the performance of the tested methods further and discusses the possible
 reasons of misclassification according to the dataset. In Section 5, we draw conclusions
 from this study.

 2. Materials and Methods
 2.1. Datasets
 This paper used the SIRI-WHU dataset [62], NWPU-RESIST45 dataset [42], UC Merced
 Land-Use dataset [63], and RSSCN7 dataset [64] for experimental analysis.
 The NWPU-RESIST45 dataset is a publicly available remote sensing image scene
 classification dataset that was created by Northwestern Polytechnical University (NWPU).
 This dataset contains 31,500 images divided into 45 scene classes. Each class consists of
 700 optical images with a size of 256 × 256 pixels. These 45 types of scenes include airplane,
 airport, baseball diamond, basketball court, beach, bridge, chaparral, church, circular
 farmland, cloud, commercial area, dense residential, desert, forest, freeway, golf course,
 ground track field, harbor, industrial area, intersection, island, lake, meadow, medium
 residential, mobile home park, mountain, overpass, palace, parking lot, railway, railway
 station, rectangular farmland, river, roundabout, runway, sea ice, ship, snowberg, sparse
 residential, stadium, storage tank, tennis court, terrace, thermal power station, and wetland.
 Figure 1 shows representative images of each class.
 The SIRI-WHU dataset was collected and produced by the research team of Wuhan
 University, which contains 12 categories of scene images including farmland, business
 quarters, ports, idle land, industrial areas, grasslands, overpasses, parking lots, ponds,
 residential areas, rivers, and water. The representative images of each class are shown in
 Figure 2. Each class in the SIRI-WHU dataset consists of 200 optical aerial images with
 200 × 200 pixels and a 2-m resolution. Dataset resources can be searched from Google
 Earth, mainly covering urban areas of China.
Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images
(10) (11) (12) (13) (14) (15) (16) (17) (18)
Remote Sens. 2021, 13, 3460 5 of 22
Remote Sens. 2021, 13, x FOR PEER REVIEW 5 of 22

 (19) (20) (21) (22) (23) (24) (25) (26) (27)

 (1) (2) (3)
 (28) (29) (30) (31) (32) (4) (33)(5) (6)
 (34) (7)
 (35) (8)
 (36) (9)

 (10) (11) (12) (13) (14) (15) (16) (17) (18)
 (37) (38) (39) (40) (41) (42) (43) (44) (45)
 Figure 1. Examples from the NWPU-RESIST45 dataset: (1) airplane, (2) airport, (3) baseball dia-
 (19)(5) beach,
 mond, (4) basketball court, (20)(6) bridge,
 (21) (7) chaparral,
 (22) (23) (24)(9) circular
 (8) church, (25) farmland,
 (26) (10)(27)
 cloud, (11) commercial area, (12) dense residential, (13) desert, (14) forest, (15) freeway, (16) golf
 course, (17) ground track field, (18) harbor, (19) industrial area, (20) intersection, (21) island, (22)
 lake, (23) meadow, (24) medium
 (28) residential,
 (29) (25) mobile
 (30) (31) home park, (26)
 (32) (33)mountain,
 (34) (27) (35)
 overpass,(36)
 (28) palace, (29) parking lot, (30) railway, (31) railway station, (32) rectangular farmland, (33) river,
 (34) roundabout, (35) runway, (36) sea ice, (37) ship, (38) snowberg, (39) sparse residential, (40) sta-
 dium, (41) storage tank, (42)
 (37) tennis(38)
 court, (43)
 (39)terrace, (44) thermal
 (40) (41) power(42)station, (45) wetland.
 (43) (44) (45)
 Figure 1. Examples from the NWPU-RESIST45 dataset: (1) airplane, (2) airport, (3) baseball dia-
 Figure 1. Examples from the NWPU-RESIST45 dataset: (1) airplane, (2) airport, (3) baseball dia-
 The SIRI-WHUmond, datasetbasketball
 was collected and produced by the research team of Wuhan (10)
 mond,(4) (4) basketballcourt,
 court,(5)(5)beach,
 beach,(6)(6)bridge,
 bridge,(7)(7)chaparral,
 chaparral,(8)(8)church,
 church,(9)
 (9)circular
 circularfarmland,
 farmland, (10)
 University, which contains
 cloud, (11) 12 categories
 commercial of scene
 area, (12) images (13)
 dense residential, including
 desert, (14)farmland, business
 forest, (15) freeway,
 cloud, (11) commercial area, (12) dense residential, (13) desert, (14) forest, (15) freeway, (16) golf
 (16) golf
 course,
 quarters, ports, idle course, (17)
 land, (17) ground
 industrial track field,
 areas, (18) harbor, (19) industrial area, (20) intersection, (21) island, (22)
 ground track field, grasslands,
 (18) harbor, (19)overpasses,
 industrial area,parking lots, ponds,
 (20) intersection, (21) island, (22)
 lake, (23) meadow, (24) medium residential, (25) mobile home park, (26) mountain, (27) overpass,
 residential areas, rivers,
 (28)
 andmeadow,
 lake,palace,
 (23) water.(24)
 The
 (29) parking
 representative
 medium
 lot, residential,
 (30) railway,
 images
 (31) (25) mobile
 railway
 of each
 home
 station,
 class
 park,
 (32)
 arefarmland,
 shown
 (26) mountain,
 rectangular
 in
 (27)(33)
 overpass,
 river,
 Figure 2. Each class (34)
 in
 (28)the SIRI-WHU
 palace, (29) parkingdataset
 lot, (30) consists
 railway, (31) of 200
 railway optical
 station, (32)aerial images
 roundabout, (35) runway, (36) sea ice, (37) ship, (38) snowberg, (39) sparse residential,river,
 rectangular with
 farmland, 200
 (33) (34)
 (40) sta-
 roundabout,
 dium, (35) runway,
 (41) storage tank, (36) sea
 (42) tennis ice, (37) ship,
 court, (43) (38)
 terrace, snowberg, (39)
 (44) thermal fromsparse residential,
 power station, (40)
 (45)Earth, stadium,
 wetland.
 × 200 pixels and a 2-m resolution. Dataset resources can be searched Google
 (41) storage tank, (42) tennis court, (43) terrace, (44) thermal power station, (45) wetland.
 mainly covering urban areas of China.
 The SIRI-WHU dataset was collected and produced by the research team of Wuhan
 University, which contains 12 categories of scene images including farmland, business
 quarters, ports, idle land, industrial areas, grasslands, overpasses, parking lots, ponds,
 residential areas, rivers, and water. The representative images of each class are shown in
 Figure 2. Each class in the SIRI-WHU dataset consists of 200 optical aerial images with 200
 × 200 pixels and a 2-m resolution. Dataset resources can be searched from Google Earth,
 mainly covering urban areas of China.
 (1) (2) (3) (4) (5) (6)

 (1) (2) (3) (4) (5) (6)
 (7) (8) (9) (10) (11) (12)
 Figure
 Figure 2. Examples
 2. Examples from from the SIRI-WHU
 the SIRI-WHU dataset:
 dataset: (1) (1)(2)farmland,
 farmland, (2) business
 business quarters, quarters,
 (3) ports, (3) (5)
 (4) idle land, ports, (4)
 industrial
 areas, (6) grasslands, (7) overpasses, (8) parking lots, (9) ponds, (10) residential areas, (11) rivers, (12)
 idle land, (5) industrial areas, (6) grasslands, (7) overpasses, (8) parking lots, (9) ponds, (10) residen- water.
 tial areas, (11) rivers, (12) water.
 The UC Merced(8)
 (7) (UCM) dataset(9) (10)
 is an aerial orthophoto (11) the national
 shot from (12)
 geologi-
 cal survey of the United States. It consists of 2100 remote sensing images
 Figure 2. Examples from the SIRI-WHU dataset: (1) farmland, (2) business quarters, (3) ports, from 21 scene
 (4)
 The UC Mercedidle
 (UCM)
 classes, asdataset
 shown inis an aerial
 Figure 3, orthophoto
 including shot
 agricultural, from the
 airplane, national
 baseball geolog-
 diamond, beach,
 land, (5) industrial areas, (6) grasslands, (7) overpasses, (8) parking lots, (9) ponds, (10) residen-
 ical survey of the United
 tial States.
 buildings,
 areas, It consists
 (11)chaparral,
 rivers, water.of
 (12)dense 2100 remote
 residential, forest, sensing images
 freeway, golf from
 course, 21 scene
 harbor, intersection,
 classes, as shown in Figure 3, including agricultural, airplane, baseball diamond, beach,sparse
 medium-density residential, mobile home park, overpass, parking lot, river, runway,
 residential,
 The UC storage
 Merced tanks,
 (UCM)and tennis
 dataset courts.
 is an aerial Each class inshot
 thefrom
 UCMthe dataset consists of
 buildings, chaparral, dense residential, forest, freeway, golforthophoto
 course, harbor, national
 intersection, geolog-
 200 survey
 ical optical of
 aerial
 the images
 United with
 States. It ×
 256 256 pixels
 consists andremote
 of 2100 a 0.3-msensing
 resolution.
 images from 21 scene
 classes, as shown in Figure 3, including agricultural, airplane, baseball diamond, beach,
 buildings, chaparral, dense residential, forest, freeway, golf course, harbor, intersection,
Cross-Domain Scene Classification Based on a Spatial Generalized Neural Architecture Search for High Spatial Resolution Remote Sensing Images
of 200
Remote Sens. 2021,optical
 13, x FOR aerial images
 PEER REVIEW with 256 × 256 pixels and a 0.3-m resolution. 6 of 22

Remote Sens. 2021, 13, 3460 medium-density residential, mobile home park, overpass, parking lot, river, runway, 6 of 22
 sparse residential, storage tanks, and tennis courts. Each class in the UCM dataset consists
 of 200 optical aerial images with 256 × 256 pixels and a 0.3-m resolution.
 (1) (2) (3) (4) (5) (6) (7)

 (1) (2) (3) (4) (5) (6) (7)
 (8) (9) (10) (11) (12) (13) (14)

 (8) (9) (10) (11) (12) (13) (14)

 (15) (16) (17) (18) (19) (20) (21)
 Figure 3. Examples from the UC Merced dataset: (1) agricultural, (2) airplane, (3) baseball diamond,
 (15)
 (4) beach, (5) buildings, (6) chaparral,(16) (17)
 (7) dense residential, (18)
 (8) forest, (9) (19)
 freeway, (10) (20)
 golf course,(21)
 (11) harbor, (12) intersection,
 Figure3.3.(13)
 Figure medium-density
 Examples
 Examples from
 fromthe
 theUC residential,
 UCMerced
 Merced dataset:(14)
 dataset: (1) mobile home
 (1)agricultural,
 agricultural,(2) park, (15)
 (2)airplane,
 airplane,(3) overpass,
 (3)baseball
 baseballdiamond,
 diamond,
 (16) parking lot, (17) river, (18) runway, (19) sparse residential, (20) storage tanks, (21) tennis courts.course,
 (4)
 (4)beach,
 beach, (5)
 (5)buildings,
 buildings,(6)
 (6)chaparral,
 chaparral,(7)
 (7)dense
 dense residential,
 residential,(8)
 (8)forest,
 forest,(9)
 (9)freeway,
 freeway, (10)
 (10)golf
 golf course,
 (11)
 (11)harbor,
 harbor,(12)
 (12)intersection,
 intersection,(13)
 (13)medium-density
 medium-densityresidential,
 residential,(14)
 (14)mobile
 mobilehome
 homepark,
 park,(15)
 (15)overpass,
 overpass,
 (16) parking lot, (17) river, (18) runway, (19) sparse residential, (20) storage tanks, (21) tennis courts.
 The RSSCN7 dataset contains 2800 remote sensing scene images, which are from courts.
 (16) parking lot, (17) river, (18) runway, (19) sparse residential, (20) storage tanks, (21) tennis

 seven typical scene categories:
 The
 The RSSCN7grassland,
 RSSCN7 datasetforest,
 dataset contains
 contains farmland,
 2800
 2800 remote
 remoteparking
 sensinglot,
 sensing sceneresidential
 scene images,
 images, which region,
 which are
 are from
 from
 seven typical scene categories: grassland, forest, farmland,
 industrial region, and river and lake. There are 400 images in each scene type, and each
 seven typical scene categories: grassland, forest, farmland, parking
 parking lot,
 lot,residential
 residential region,
 region,
 industrial
 image has a size of 400 × 400region,
 industrial region,and
 pixels. and river
 It isriver and
 andlake.
 important lake.toThere
 noteare
 There are 400 images
 400the
 that images in
 sampleineach
 each scene
 scene type,
 images type, and
 and each
 in each each
 image
 image has a size of 400 × 400 pixels. It is important to note that the sample imagesin
 has a size of 400 × 400 pixels. It is important to note that the sample images ineach
 class are sampled onclass
 four different scales with 100scalesimages per scale with imaging each
 differentdifferent
 classare
 aresampled
 sampledon onfour
 fourdifferent
 different scaleswith with100
 100images
 imagesper perscale
 scalewith
 with differentimaging
 imaging
 angles. The RSSCN7 dataset
 angles.
 angles. Theis
 The a challenging
 RSSCN7 dataset isisaaand
 dataset representative
 challenging
 challenging scene classification
 andrepresentative
 and representative scene
 scene da-dataset
 classification
 classification da-
 taset due to the widetaset
 diversity
 due due
 to the ofwide
 to the
 wide the images,
 diversity of which
 diversity of
 thethe are captured
 images,
 images, whichare
 which areunder
 captured
 captured changing
 under changingseasons,
 changing seasons,
 seasons,
 varying weathers, andvarying
 varying weathers,
 on and
 weathers,
 sampled andsampled
 sampled
 different on
 ondifferent
 scales different
 [64], as scales
 scales
 shown [64], as
 [64],inas shown
 shown in
 Figure in4.Figure
 Figure 4. 4.

 (1) (2) (3) (4) (5) (6) (7)
 (1) (2)Figure 4. Examples
 (3) from the RSSCN7
 (4) dataset: (1)(5) (6) (3) farmland,(7)
 grassland, (2) forest, (4) parking lot,
 (5) residential region, (6) industrial region, (7) river and lake.
 Figure
 Figure 4. 4. Examples
 Examples fromfrom the RSSCN7
 the RSSCN7 dataset: dataset: (1) grassland,
 (1) grassland, (2) forest, (3)(2) forest,(4)
 farmland, (3)parking
 farmland,
 lot, (5)(4) parkingregion,
 residential lot,
 (6)
 (5)industrial region,
 residential (7) river
 region, (6)and lake.
 industrial
 2.2. Methods region, (7) river and lake.
 2.2. A spatial generalization neural architecture search framework was proposed in this
 Methods
 2.2. Methods paper for cross-domain scene classification of HSR remote sensing images. As shown in
 A spatial generalization neural architecture search framework was proposed in this
 Figure 5, the process of cross-domain scene classification of this paper can be divided into
 A spatial generalization neural architecture
 paper for cross-domain search framework
 scene classification of HSR remote was proposed
 sensing images.in Asthis
 shown in
 three steps. (1) The architecture search space is designed. (2) The designed search space
 paper for cross-domainFigure 5, theclassification
 scene process of cross-domain
 of HSR scene classification
 remote sensing of this paper
 images. As can be in
 shown divided
 searched network for HSR images dataset. The HSR image dataset is used as a search
 into three
 Figure 5, the processdataset steps.
 of cross-domain (1) The architecture search space is designed. (2) The designed search
 and input intoscene classification
 the designed of this
 search space paperfor
 to search can be divided
 networks in thisinto
 step. The
 space searched network for HSR images dataset. The HSR image dataset is used as a
 three steps. (1) The model
 architecture
 search dataset search space
 is continuously optimized
 and input is through
 into the designed. (2) The
 the performance
 designed search designed
 space to searchsearch
 evaluation feedback space
 for networks
 until the
 in this
 searched network for final model
 HSR is obtained. (3) The network obtained by the search space is used for cross-
 step. The images dataset. The
 model is continuously HSR image
 optimized through dataset is used evaluation
 the performance as a search feedback
 domain scene classification to verify the generalization ability. An HSR image dataset is
 dataset and input into the
 until thedesigned
 final modelsearch space
 is obtained. (3)to
 Thesearch
 network forobtained
 networks insearch
 by the this step.
 space The
 is used for
 used as the
 cross-domain training
 scene dataset, and
 classification another HSR dataset
 to verify the generalizationdifferent from
 ability. the
 An HSR training
 image set is
 dataset
 model is continuouslyused optimized
 as as
 thethe through
 testing dataset. the performance
 Finally, throughHSR evaluation
 cross-domain feedback until
 scene classification, the
 the cate-
 is used training dataset, and another dataset different from the training set is
 final model is obtained.
 gory (3) The
 labels of network
 different obtained
 images will be by the search space is used for cross-
 output.
 used as the testing dataset. Finally, through cross-domain scene classification, the category
 domain scene classification to verify
 labels of different the generalization
 images will be output. ability. An HSR image dataset is
 used as the training dataset, and another HSR dataset different from the training set is
 used as the testing dataset. Finally, through cross-domain scene classification, the cate-
 gory labels of different images will be output.
Remote Sens.
Remote Sens. 2021, 13, x3460
 2021, 13, FOR PEER REVIEW 7 7of
 of 22
 22

 Figure 5. Cross-domain
 Figure scene
 5. Cross-domain classification
 scene process
 classification based
 process on the
 based SGNAS.
 on the SGNAS.

 2.2.1. Search
 2.2.1. Search Space
 Space
 A low-dimensional
 A low-dimensional and and simple
 simple search
 search space
 space was
 was designed
 designed inin this
 this paper.
 paper. The
 The archi-
 archi-
 tecture suitable for cross-domain scene classification of HRS remote sensing images was
 tecture suitable for cross-domain scene classification of HRS remote sensing images was
 searched based on the designed search space. The search space combines the generalization
 searched based on the designed search space. The search space combines the generaliza-
 of manually designed networks, which can generalize them in various networks and tasks.
 tion of manually designed networks, which can generalize them in various networks and
 The purpose of this study was to optimize the problems existing in the manually designed
 tasks. The purpose of this study was to optimize the problems existing in the manually
 network and traditional NAS. After designing a simple and low-dimensional search space,
 designed network and traditional NAS. After designing a simple and low-dimensional
 the generalization ability of the search space was evaluated. He et al. verified the search
 search space, the generalization ability of the search space was evaluated. He et al. verified
 space of Regnet [56] and found that the search space had no signs of overfitting under the
 the search space of Regnet [56] and found that the search space had no signs of overfitting
 conditions at higher flops, higher epochs, with 5-stage networks, and with various block
 under the conditions at higher flops, higher epochs, with 5-stage networks, and with var-
 types. These phenomena indicate that the search space can be extended to new settings.
 ious block types. These phenomena indicate that the search space can be extended to new
 Liu et al. also used Regnet as a comparison network for image classification in the field of
 settings. Liu et al. also used Regnet as a comparison network for image classification in
 computer vision [65]. Inspired by the above research, a network search space that combines
 the
 the field of computer
 advantages visiondesigned
 of manual [65]. Inspired by the
 network andabove research,
 traditional NAS a network searchin
 was designed space
 this
 that combines the advantages of manual designed network and traditional NAS
 research. The search space also retained the semiautomatic performance of traditional NAS, was de-
 signed
 which canin this research.
 search a goodThe searchfor
 network space also retained
 the specific the semiautomatic
 architecture automatically.performance
 The essence ofof
 traditional NAS, which can search a good network for the specific
 SGNAS is a semiautomatic NAS combined with the generalization ability of the architecture automati-
 manual
 cally.
 designThe essenceSection
 network. of SGNAS is a be
 2.2.1 can semiautomatic NASparts:
 divided into two combined with of
 (A) Design thethe
 generalization
 search space
 ability of the manual design network.
 and (B) search space quality detection. Section 2.2.1 can be divided into two parts: (A) De-
 sign of the search space and (B) search space quality detection.
 A. Design of the Search Space
 A. Design of the space
 The search SearchofSpace
 SGNAS will define a series of basic network operations (convolu-
 tion,The
 fullysearch spacelayer,
 connected of SGNAS
 averagewillpooling,
 define aand
 series of basic
 residual network block)
 bottleneck operations (convo-
 and connect
 lution, fully connected
 these operations to formlayer, averagenetwork.
 an effective pooling, and
 The residual
 design ofbottleneck
 the searchblock) and
 space in connect
 this study
 these operations
 is a process to form
 of gradual an effective network.
 simplification Thesearch
 of the initial designspace
 of thewithout
 search space in this[56].
 constraints studyA
 simple
 is schematic
 a process diagram
 of gradual of the design
 simplification concept
 of the initialissearch
 shown in Figure
 space 6 [56].
 without constraints [56]. A
 simple Figure 6a shows
 schematic that of
 diagram thethe
 search
 designspace was simplified
 concept is shown ingradually from A to B to C in
 Figure 6 [56].
 the design of the search space in this paper. The error distribution was strictly from A to
 B to C, as is shown in Figure 6b. Each design step was used to search for a simpler and
 more efficient model than before. The input was the initial search space without constraints
 in this paper, while the model output was simpler, had lower computational cost, and
 better performance than the previous model. The basic network design of the initial search
 space was also very simple (as shown in Figure 7 [56]). The search space was composed
 of the body that was used to perform the calculation, stem, and full connection layer
 (head) for predicting output types (as shown in Figure 7a), and we set the input picture
 to 224 × 224 × 3 in this study. The stem is a convolution layer, where the convolution
 kernel was 3 × 3, the number of convolution kernels was 32, and the value of strides was 2.
Figure 6. The design process of the architecture search space in this paper.

 Remote Sens. 2021, 13, 3460 Figure 6a shows that the search space was simplified gradually 8 of 22 fro
 the design of the search space in this paper. The error distribution was
 B to C, as is shown in Figure 6b. Each design step was used to search
 The head was a fully connected layer, which was used to output n categories. The body
Sens. 2021, 13, x FOR PEER REVIEW
 more efficient model
 was composed of four stages.than before.
 From stage The4,input
 1 to stage was gradually
 the resolution the initial
 halvedsearch
 (as sp
 shown in Figure 7b). However, the number of characteristic
 straints in this paper, while the model output was simpler, had lower cographs output by each stage,
 which are ω1 , ω2 , ω3 , and ω4 , needs to be searched as hype-parameters in the search
 and better
 space. performance
 Each stage was composed than the previous
 of di identical blocks (asmodel. The basic
 shown in Figure 7c). network de
 search space was also very simple (as shown in Figure 7 [56]). The searc
 posed of the body that was used to perform the calculation, stem, and ful
 (head) for predicting output types (as shown in Figure 7a), and we set th
 224 × 224 × 3 in this study. The stem is a convolution layer, where the co
 was 3 × 3, the number of convolution kernels was 32, and the value of s
 head was a fully connected layer, which was used to output n categori
 composed of four stages. From stage 1 to stage 4, the resolution grad
 shown in Figure 7b). However, the number of characteristic graphs outp
 which are , , , and , needs to be searched as hype-paramet
 space. Each stage was composed of identical blocks (as shown in Fig
 Figure 6. The design process of the architecture search space in this paper.

 Figure 6. The design process of the architecture search space in this paper.

 Figure 6a shows that the search space was simplified gradually from
 the design of the search space in this paper. The error distribution was s
 B to C, as is shown in Figure 6b. Each design step was used to search fo
 more efficient model than before. The input was the initial search spa
 straints in this paper, while the model output was simpler, had lower com
 and better performance than the previous model. The basic network des
 search space was also very simple (as shown in Figure 7 [56]). The search
 posed of the body that was used to perform the calculation, stem, and full
 (head) for predicting output types (as shown in Figure 7a), and we set the
 224 × 224 × 3 in this study. The stem is a convolution layer, where the con
 was 3 × 3, the number of convolution kernels was 32, and the value of st
 head was
 Figure a fully
 7. The connected
 basic network layer,
 design of the which
 initial search spacewas usedinto
 constructed thisoutput
 paper. n categorie
 Figure
 composed 7. The
 of four stages. From stage 1 to stage 4, the resolution in
 basic network design of the initial search space constructed this
 gradu
 The block used in this paper was the residual bottleneck block [56,66], which is the
 shown
 block in Figure
 of the halved7b). widthHowever,
 and halved the height number of characteristic
 of the output feature map (as shown graphs in outpu
 The block
 The main used
 branch in
 of this
 the paper
 block is a 1 was
 which are , , , and , needs to be searched as hype-paramete
 Figure 8). × 1 the residual
 convolution (including bottleneck
 BN and block
 ReLU), a [56
 3 × 3
 blockEach group convolution
 of the halved (including BN
 width and halved and ReLU), and a 1 × 1 convolution (including
 space.
 BN). On the stage
 branch was of the composed
 shortcut, whenof height
 stride =identical
 of the output
 blocks
 1, no processing (as feature
 shown
 is performed.
 map (a
 Whenin Figu
 8).stride
 The=main branch of the
 2, it is down-sampled block
 through a 1is× a1 1 × 1 convolution
 convolution (including BN) (including
 [56]. The BN a
 group
 number convolution
 of blocks also needs (including BN Three
 to be searched. and parameters
 ReLU), and need a to1be×searched
 1 convolution
 in the (in
 residual bottleneck block: the number of channels of block (ω1 ), cell group width (the
 thenumber
 branch of the repeats
 of horizontal shortcut, when
 of block, stride
 gi ), and = 1, nothatprocessing
 the parameter is performed
 is used to determine the
 it is down-sampled
 number of characteristic through
 graphs inside a 1the×block
 1 convolution
 (bottleneck ratio(including
 bi ). Generally,BN)there[56].
 are The
 four stages in the network, while the parameters that need to be determined in each stage
 also needs to be searched. Three parameters need to be searched in the re
 are di and ωi , bi , gi . Therefore, the optimal setting of 16 parameters is needed to design the
 block:
 search the
 space.number of channels of block ( ), cell group width (the num
 repeats of block, ), and the parameter that is used to determine
Remote Sens. 2021, 13, x FOR PEER REVIEW 9 of 22

 characteristic graphs inside the block (bottleneck ratio ). Generally, there are four stages
Remote Sens. 2021, 13, 3460 in the network, while the parameters that need to be determined in each stage are 9and of 22
 , , . Therefore, the optimal setting of 16 parameters is needed to design the search
 space.

 Figure
 Figure8.8.The
 Thestructure
 structureof
 ofresidual
 residualbottleneck
 bottleneckblock.
 block.(a)
 (a)Case
 Caseof
 ofstride
 stride==1,1,(b)
 (b)Case
 Caseof
 ofstride
 stride==2.2.

 Followingthe
 Following thesettings
 settings of of [56], this study limited the initial values of ω i ,,d 
 [56], this , i , and
 i, b , andgi
 to ω ≤ 1024, d ≤
 toi ≤ 1024,i ≤ 16, 16, b ∈
 i ∈ {1,2,4},
 {1,2,4}, g i ∈ ∈ {1, 2, …, 32}. The network key parameters
 {1, 2, . . . , 32}. The network key parameters were
 extracted
 were from from
 extracted the above rangerange
 the above by log-uniform
 by log-uniform sampling, and the
 sampling, andquality of search
 the quality space
 of search
 was judged
 space by EDF
 was judged byfunction [56]. After
 EDF function [56].that,
 After the search
 that, the space
 searchofspace
 SGNAS was continuously
 of SGNAS was con-
 optimizedoptimized
 tinuously and updated andby updating
 updated bythe factors such
 updating as the sharing
 the factors such as bottleneck
 the sharingrate, sharing
 bottleneck
 group width, increasing the number of channels, and increasing
 rate, sharing group width, increasing the number of channels, and increasing the number the number of blocks. The
 accuracy of the model will not be lost due to the change in the above
 of blocks. The accuracy of the model will not be lost due to the change in the above factors, factors, as can be
 proven by determining the quality of the search space. In contrast,
 as can be proven by determining the quality of the search space. In contrast, the search the search range of the
 searchofspace
 range can be greatly
 the search space can reduced, and the
 be greatly search and
 reduced, efficiency can beefficiency
 the search improvedcan by sharing
 be im-
 the group width. The performance of the architecture can be
 proved by sharing the group width. The performance of the architecture can be improved improved by increasing the
 number of channels and blocks. On the basis of the previous
 by increasing the number of channels and blocks. On the basis of the previous optimiza-optimization and the results
 obtained
 tion and theby results
 setting obtained
 AnyNetB,byAnyNetC, AnyNetD,
 setting AnyNetB, and AnyNetE
 AnyNetC, in [56],
 AnyNetD, it can
 and be found
 AnyNetE in
 that the number of channels of the block (ω 0 ) has a linear relationship
 [56], it can be found that the number of channels of the block ( ) has a linear relationship with the index of
 blocks in the search space of the SGNAS. The linear function is:
 with the index of blocks in the search space of the SGNAS. The linear function is:
 µ j ==ω 0 ++ω s ∗∗j( (0
 0
 < 256, ∈ [1.5, 3], ∈ {1, 2, 4} and ∈{1, 2, …, 32}, which follows the setting of [56].
 The search space will be simplified by using the human-the-loop method after the basic
 setting of the search space is completed. The simplification of the search space aims to
 obtain a search space with low computational cost and low number of epochs, in order to
Remote Sens. 2021, 13, 3460 improve the computational efficiency of the search space. Based on the original low-di-10 of 22
 mensional and simple search space, this study completed the generalization setting based
 on the design of Regnet [56]. The finished SGNAS is a semiautomatic network architec-
 ture.
 space, this study completed the generalization setting based on the design of Regnet [56].
 The finished SGNAS is a semiautomatic network architecture.
 B. Search Space Quality Detection
 The search
 B. Search Spacespace
 Qualitydesigned
 Detectionin this paper was an enormous space containing many
 models.The search space designed infrom
 We can sample the model this apaper
 search
 wasspace to generatespace
 an enormous a model distribution
 containing many
 and then use classical statistical tools to evaluate the search space. In this
 models. We can sample the model from a search space to generate a model distributionstudy, an error
 empirical
 and thendistribution
 use classicalfunction (EDF)
 statistical tools [67] was used
 to evaluate thetosearch
 detectspace.
 the quality
 In this of the an
 study, design
 error
 space. When n models are sampled in the search space and the error is , EDF
 empirical distribution function (EDF) [67] was used to detect the quality of the design is:
 space. When n models are sampled in the search space and the error is mi , EDF is:
 1
 ( ) = 1 
Remote Sens. 2021, 13, 3460 11 of 22

 search space was optimized during the design of the search space in this study, thus
 improving the search efficiency of the framework, the search strategy adopted in this study
 was a simple random search strategy. Random search is the simplest search strategy, which
 randomly selects an effective candidate architecture from the search space and does not
 involve a learning model. However, it has been proven to be very useful in hype-parameter
 search [54].

 B. Performance Estimation
 The evaluation step is shown in Figure 9 and requires an effective performance
 estimation method. The simplest performance estimation method of training from scratch
 is used to measure the accuracy of the searched model through experiments [54], which
 is used to obtain feedback to optimize the search algorithm. The proposed framework
 uses this feedback to retrain the network obtained from the previous search, in order to
 optimize the model parameters of the model searched previously. Then, a model that is
 more suitable for cross-domain scene classification can be obtained. Although this method
 is slow, the search space of the SGNAS is carried out in a low-computation, low-epoch
 regime, which greatly improves the computing speed.

 2.2.3. Cross-Domain Scene Classification
 The SGNAS was used to search for the image features of the HSR remote sensing image
 dataset, in order to obtain the final model suitable for cross-domain scene classification
 of the searched dataset. The searched model will be migrated to different datasets for
 cross-domain scene classification to obtain the final classification results, which aims to
 verify the spatial generalization of the model obtained by the proposed framework. Specific
 experimental settings and experimental results are shown below.

 3. Results
 Cross-domain scene classification experiments between different datasets were used
 to verify the generalization of the framework and confirm the complexity of the SGNAS
 in order to avoid overfitting. The quantitative and computational performance of the
 searched network were evaluated by overall accuracy (OA), confusion matrix, and inference
 time (Time_infer).

 3.1. Experimental Setup
 3.1.1. Implementation Details
 The experiments were implemented using PyTorch. The cross entropy loss function
 (CrossEntropyLoss) was used as the loss function. The learning rate was initially set to 1e-5
 with a weight decay of 5e-6. The mini-batch size was set to eight, and the normalization
 parameter was set to 5e-5. All models were trained on four NVIDIA RTX2080 GPUs.
 To verify the spatial generalization ability of the proposed model, we established three
 cross-domain scenarios termed as UCM→RSSCN7, SIRI-WIIU→RSSCN7, and NWPU-
 RESIST45→RSSCN7, referring to source domain→target domain for data analysis. As the
 training set and test set of this experiment came from two different datasets, there were large
 differences in image types, scales, and contents. To match the seven types in the RSSCN7
 dataset, for the UCM dataset, this experiment selected the corresponding seven similar
 types for the experiment: golf course, agricultural, storage tanks, river, forest, density
 residential, and parking lot. For the SIRI-WHU dataset, industrial areas, farmland, parking
 lots, residential areas and rivers were selected to correspond to the five types of industrial
 region, farmland, parking lot, residential region, and river and lake in the RSSCN7 dataset.
 Seven similar types were selected from the NWPU-RESIST45 dataset for the experiments:
 dense residential, forest, golf course, industrial area, parking lot, rectangular farmland,
 and river. The experiments of AlexNet, Vgg16, ResNet50 [68], and ENAS [69] on the same
 cross-domain scenario settings were used for comparative validation.
Remote Sens. 2021, 13, 3460 12 of 22

 3.1.2. Evaluation Metrics
 The OA, inference time, and confusion matrix are employed as the evaluation standard
 Remote Sens. 2021, 13, x FOR PEER REVIEW
 of the cross-domain scene classification. The OA is defined as the number of correctly 12 of 22
 classified images divided by the total number of images. The inference time implies the
 inference efficiency of the trained model to make predictions against previously unseen
 data, andlots,
 parking is used as an evaluation
 residential metricwere
 areas and rivers to validate
 selectedthe
 to efficiency
 correspondoftodifferent
 the five models
 types of
 in the experiment. The confusion matrix is an informative table
 industrial region, farmland, parking lot, residential region, and river and used for analyzing
 lake inthethe
 confusions between different scene classes. It is obtained by counting
 RSSCN7 dataset. Seven similar types were selected from the NWPU-RESIST45 dataset forthe correct and
 incorrect classifications
 the experiments: of the test images
 dense residential, in each
 forest, golf classindustrial
 course, and accumulating the lot,
 area, parking results in
 rectan-
 agular
 table.farmland, and river. The experiments of AlexNet, Vgg16, ResNet50 [68], and ENAS
 [69] on the same cross-domain scenario settings were used for comparative validation.
 3.2. Experiment 1: UCM→RSSCN7
 3.1.2.The UCM dataset
 Evaluation was used as the training set, and the RSSCN7 dataset was used as
 Metrics
 the test set in this section to conduct the cross-domain scene classification experiment. The
 The OA, inference time, and confusion matrix are employed as the evaluation stand-
 accuracy and the inference time comparison between SGNAS and other models is shown in
 ard of the cross-domain scene classification. The OA is defined as the number of correctly
 Table 1, and the accuracy of each scene is shown in Table 2. Figure 10 shows the confusion
 classified
 matrix images
 of the divided results.
 classification by the total number of images. The inference time implies the
 inference efficiency of the trained model to make predictions against previously unseen
 data,1.and
 Table is used
 Overall as anand
 accuracy evaluation
 inference metric to validate the
 time of cross-domain efficiency
 scene of different
 classification models
 of different modelsin
 the experiment.
 on UCM→RSSCN7. The confusion matrix is an informative table used for analyzing the con-
 fusions between different scene classes. It is obtained by counting the correct and incorrect
 Models
 classifications OA (%)
 of the test images in each class and accumulating theTime_infer
 results in a(s)table.
 Vgg16 43.79 23
 AlexNet
 3.2. Experiment 1: UCM→RSSCN7 44.00 12
 ResNet50 46.71 9
 The UCM ENAS
 dataset was used as the training
 40.82
 set, and the RSSCN7 dataset
 46
 was used as
 the test setSGNAS
 in this section to conduct the cross-domain
 59.25 scene classification experiment.
 8 The
 accuracy and the inference time comparison between SGNAS and other models is shown
 in Table 1, and the accuracy of each scene is shown in Table 2. Figure 10 shows the confu-
 Table 2. Accuracy
 sion matrix of classification
 of the various scenesresults.
 in cross-domain scene classification of the UCM and RSSCN7
 datasets of SGNAS.
 Table 1. Overall accuracy and inference time of cross-domain scene classification of different models
 on UCM→RSSCN7. Class Accuracy (%)
 River and lake 82.75
 Models
 Residential region OA (%) 66.75Time_infer (s)
 Vgg16 Parking lot 43.79 21.25 23
 Industrial region
 AlexNet 44.00 39.50 12
 Grassland 77.50
 ResNet50 46.71 9
 Forest 63.75
 ENASFarmland 40.82 63.25 46
 SGNAS 59.25 8

 Figure 10.Confusion
 Figure10. Confusionmatrix
 matrixof
 ofthe
 thecross-domain
 cross-domainscene
 sceneclassification
 classificationresults
 resultsfor
 forUCM →RSSCN7.
 UCM→RSSCN7.

 Table 1 shows that the experimental results of the model searched by the SGNAS
 were superior to those of the models used for comparison (i.e., the manually designed
 Vgg16, AlexNet and ResNet50) and the network searched by ENAS. From Table 2 and
 Figure 10, it can be seen that the overall classification accuracy of river and lake was the
Remote Sens. 2021, 13, 3460 13 of 22

 Table 1 shows that the experimental results of the model searched by the SGNAS were
 superior to those of the models used for comparison (i.e., the manually designed Vgg16,
 AlexNet and ResNet50) and the network searched by ENAS. From Table 2 and Figure 10,
 it can be seen that the overall classification accuracy of river and lake was the highest,
 with the classification accuracy of 82.75%, while the parking lot and industrial regions
 had lower accuracy. The classification accuracy of the parking lot was 21.25%, and that
 of the industrial region was 39.50%. However, among the seven scenes, the classification
 accuracies of five scenes were over 60%. As can be seen in the confusion matrix, there
 was some confusion between certain scenes. For instance, some scenes belonging to the
 parking lot were classified as grassland. The inference time used by this framework was
 less than that of the other models used for comparison, as shown in Table 1. This indicates
 that the proposed SGNAS framework is a promising avenue for efficient cross-domain
 scene classification.

 3.3. Experiment 2: SIRI-WIIU→RSSCN7
 This experiment uses the SIRI-WHU dataset as the training set and the RSSCN7 dataset
 as the test set to conduct the cross-domain scene classification experiment. The accuracy
 and the inference time comparison between SGNAS and other models is shown in Table 3,
 and the accuracy of each scene is shown in Table 4. Figure 11 shows the confusion matrix
 of the classification results.

 Table 3. Overall accuracy and inference time of the cross-domain scene classification of different
 models for SIRI-WHU→RSSCN7.

 Models OA (%) Time_infer (s)
 Vgg16 40.30 37
 AlexNet 45.25 11
 ResNet50 42.40 142
Remote Sens. 2021, 13, x FOR PEER REVIEW ENAS 46.95 137 14 of 22
 SGNAS 53.95 40

 Table 4.
 Table Accuracyof
 4. Accuracy ofvarious
 variousscenes
 scenesininthe
 thecross-domain
 cross-domainscene
 sceneclassification
 classificationofofthe
 theSIRI-WHU
 SIRI-WHUand
 and
 RSSCN7
 RSSCN7datasets
 datasetsof
 ofthe
 theSGNAS.
 SGNAS.

 Class
 Class Accuracy
 Accuracy (%)
 (%)
 Farmland
 Farmland 79.75
 79.75
 Industrial
 Industrialregion
 region 8.25
 8.25
 Parking lot
 Parking lot 49.75
 49.75
 Residential region
 Residential region 65.25
 65.25
 River and lake 66.75
 River and lake 66.75

 Figure11.
 Figure Confusionmatrix
 11.Confusion matrixof
 ofcross-domain
 cross-domainscene
 sceneclassification
 classificationresults
 resultsin SIRI-WHU→RSSCN7.
 inSIRI-WHU→RSSCN7.

 3.4. Experiment 3: NWPU-RESIST45→RSSCN7
 The NWPU-RESIST45 dataset was used as the training set and the RSSCN7 dataset
 was used as the test set to conduct the cross-domain scene classification experiment in this
You can also read